New User Tag Recommendation Research Based On Tensor Decomposition

Due to the rapid growth of social network information, it is difficult for users to obtain accurate information quickly, social tag recommendation technology is designed to recommend tags and resources to the users actively through the users’historical behaviors, allowing users to find valuable information from the vast amount of information, and meanwhile letting valuable information be shared. Recently tensor decomposition models are used in social tag recommendation system to obtain high-quality recommendation results. This thesis mainly focuses on social tag tensor decomposition models and effective solutions based on these models for new user recommendation.Based on traditional tensor decomposition models Tucker and ParaFac, the corresponding new user tag recommendation algorithms are proposed to deal the new user tag recommendation problem, by updating original tensor decomposition models. Then experimental simulation are conducted for the two proposed new user tensor prediction algorithms on three real data sets, such as Last.fm, Bibsonomy and Movielens, experimental results show that these two new user tag recommendation algorithms have greatly improved the precision and recall rate of recommendation results.High-dimensional Tensor of the traditional tensor decomposition models Tucker and ParaFac has led to high levels of time complexity to the algorithm, and due to the sparsity of tag datas, the traditional tensor decomposition tag recommendation algorithms would overfitting and reduce the accuracy of the recommendation algorithm. In order to improve the algorithm accuracy, this thesis uses the low-order tensor decomposition (LOTD) which is more accord with sparse data statistical properties, and proposes a new user prediction algorithm by updating the original low-order tensor decomposition model. Similarly, experiment comparative analysis of three new user tag recommendation algorithms are conducted on three real data sets Last.fm, Bibsonomy and Movielens. Experimental results show that the new user tag recommendation algorithm based on low-order tensor decompositon model in this paper has higher recommendation quality, and more effective to sparse data set.